Damien Charlotin's hallucination database catalogues 1,353+ court decisions worldwide that have called out AI-generated fabrications in legal filings, with the pace accelerating sharply in 2026. ABA Model Rule 1.1 makes the attorney — not the model — responsible. A federal judge in Oregon sanctioned two lawyers $110,000 in 2026 over 23 fabricated citations and 8 invented quotations: the largest AI-hallucination sanction in US legal history. Q1 2026 alone saw US courts impose over $145,000in AI-hallucination sanctions across multiple cases. VeriSwarm produces the audit evidence that proves a legal AI agent's output was grounded, reviewed, and approved — at the point of generation, not after a Rule 11 motion. Works across LangChain, CrewAI, AutoGen, your firm's in-house drafting stack, or any other framework you use.
Five ethics duties intersect with every AI-assisted filing. We mapped each one to the VeriSwarm capability that produces the evidence the court asks for — and the API endpoint that emits it.
“Lawyers must understand the benefits and risks of relevant technology. The attorney — not the model — is responsible for what is signed and filed.”
Every AI-assisted draft step recorded with prompt, model, output, and reviewer action in an immutable chain. The attorney's review is itself a Vault event with timestamp and approver identity.
Audit evidence: Per-draft provenance record showing every AI exchange and human review step. The lawyer's competence with the technology is documented, not asserted.
POST /v1/workflows/execute“Lawyers shall not knowingly make false statements of fact or law to a tribunal. Citation verification is a duty, not an option.”
Citations routed through multiple LLMs with majority-vote consensus on existence and accuracy. Grounding source-attribution recorded for every quoted passage. Hallucinated citations fail the verification gate before the draft is touchable.
Audit evidence: Per-citation verification record with which model said what and the corresponding authority. Rule 11 motion gets answered with a JSON export.
POST /v1/suite/guard/scan“Privileged content must not leak through AI tools. Client names, matter numbers, and confidential facts cannot become training data or model logs.”
Recursive token replacement for client names, matter numbers, and privileged identifiers before the LLM call. Guard Proxy transparently intercepts MCP tool traffic for in-flight enforcement.
Audit evidence: Per-session record showing what was tokenized, what reached the model, and what was rehydrated for the lawyer. Privilege is preserved cryptographically.
POST /v1/suite/guard/pii/tokenize“Lawyers responsible for the conduct of non-lawyer assistants — including their AI tools.”
Per-agent trust score across identity, risk, reliability, autonomy. Signed credentials with audience binding and 1-hour TTL. Supervising attorney can pause, gate, or revoke an agent's authority at any time.
Audit evidence: Per-agent trust score timeline, supervisor approval records, kill-switch events. Supervisory record meets Rule 5.3 documentation expectations.
POST /v1/decisions/check“Federal Rules of Civil Procedure Rule 11 — every filing must be certified as non-frivolous, grounded in fact, and warranted by existing law or non-frivolous argument for extension.”
Mandatory human-review gate in the drafting workflow. Approver identity, timestamp, and the diff approved all written as Vault events. Chain verification proves the review actually happened.
Audit evidence: Per-filing Vault chain export — every AI input, every citation check, every reviewer approval. The sanctions motion arrives; the chain is the response.
GET /v1/suite/vault/verifyThe base rate of detection is now daily; the sanctions are getting larger.
Source: Damien Charlotin AI Hallucination Cases Database; Sterne Kessler 2025 review; Reason / Volokh Conspiracy 2026-04-06; Fortune 2026-05-16.
The detection apparatus is now permanent. The question is what evidence you have when it finds you.
A federal judge in Oregon sanctioned two lawyers $110,000 in 2026 after they submitted a brief containing 23 fabricated citations and 8 invented quotations. Q1 2026 alone saw US courts impose over $145,000 in sanctions for AI hallucinations in legal filings.
Fortune, 2026-05-16; Sterne Kessler 2025 review of sanctions
The most comprehensive public tracker of AI hallucination cases in law. Searchable by country, party, AI tool, and outcome. The database has been cited in several court decisions itself, and its companion PelAIkan automated reference checker now operates against new filings.
damiencharlotin.com/hallucinations, accessed 2026-05-19
Per Reason / Volokh Conspiracy tracking, 17 separate US court decisions noted suspected AI hallucinations in legal filings on a single day. Daily detection volume is now routine. The question is no longer "if your tool will hallucinate" — it's "what evidence will you produce when the Rule 11 motion lands."
Reason.com / Volokh Conspiracy, 2026-04-06
Two attestations behind every AI-assisted filing: professional-responsibility ethics (ABA Model Rules + state-bar opinions) and pre-filing sanction defense (FRCP Rule 11 + Oregon $110K precedent).
Rules 1.1 / 1.6 / 3.3 / 5.1-5.3 mapped to Cortex docs, Guard PII tokenization, hallucination guardrails, and per-state Cedar policy variants.
See attestation map →Citation-check workflow, grounding evidence, pre-filing human_review — mapped to the Oregon $110K (2026) precedent's sanctions standard.
See attestation map →Pick the ethics duty you've been asked about most recently. We'll show you the audit-evidence artifact for that duty, generated against live agent traffic in 30 minutes. No demo data — you bring the question.
Sources: Damien Charlotin AI Hallucination Cases Database (accessed 2026-05-19); ABA litigation news (2025); Reason/Volokh Conspiracy (2026-04-06); Fortune (2026-05-16); Sterne Kessler 2025 review of sanctions.